Zgodnje zaznavanje biotskega in abiotskega stresa rastlin krompirja (Solanum tuberosum L.) z uporabo spektralnega slikanja in umetne inteligence

Lapajne, Janez (avtor), Žibrat, Uroš (mentor), Širca, Saša (komentor)

Zgodnje zaznavanje zdravstvemega stanja rastlin je bistveno za ohranjanje njihovega zdravja in količine pridelkov ter spodbujanje trajnostnih načinov kmetovanja. Pravočasno prepoznavanje dejavnikov stresa je mogoče z uporabo spektralnega slikanja, ki predstavlja neinvazivno metodo za spremljanje odzivov rastlin na stresne dejavnike. Disertacija preučuje združitev spektralnega slikanja in strojnega učenja za zaznavanje zdravstvenega stanja krompirja, pri čemer raziskave temeljijo na multispektralnih in hiperspektralnih podatkih laboratorijskih in poljskih poskusov. Za izboljšanje natančnosti in interpretacije zaznavanja stresa smo uporabili napredne pristope modeliranja, vključno z globokimi nevronskimi mrežami, ojačevalnimi modeli in tehnikami razložljivosti, ter metodami za določanje ključnih spektralnih valovnih dolžin, povezanih s specifičnimi stresnimi odzivi. Rezultati kažejo, da modeli strojnega učenja zanesljivo razločujejo med zdravimi in rastlinami podvrženimi stresu, še preden znaki stresa postanejo vidni s prostim očesom. Modeli dosegajo F1-točnost do 0,95 pri zaznavanju vodnega stresa, do 100 % zanesljivost pri prepoznavanju rastlin in gomoljev okuženih z ogorčicami in F1-točnost 0,826 pri zaznavanju črne listne pegavosti krompirja. Kljub temu ostajajo izzivi pri razlikovanju med biotskimi in abiotskimi stresorji, zlasti pri sočasni prisotnosti več dejavnikov. Analiza je identificirala ključne spektralne pasove za zaznavanje stresa: 475–580 nm, 660–730 nm, 940–970 nm, 1420–1510 nm, 1875–2040 nm in 2350–2480 nm za vodni stres. Bližnje-infrardeči del spektra je bil najpomembnejši za zgodnje zaznavanje okužb s črno listo pegavostjo krompirja, takoj pred modrim in zelenim kanalom. Spektralno območje 1000–1200 nm je bilo povezano z notranjostjo gomoljev, napadenih z ogorčicami, medtem ko so pasovi 1500–1600 nm, 1850–2000 nm in 2300–2450 nm ustrezali njihovi zunanjosti. Ugotovitve raziskave pomembno prispevajo k razumevanju rastlinske fiziologije in napredku preciznega kmetijstva. Združitev spektralnega slikanja in umetne inteligence omogoča razvoj inovativnih orodij za zgodnje odkrivanje stresa, kar prispeva k boljši prilagoditvi na podnebne spremembe ter zaščiti pridelkov pred škodljivci in boleznimi.

• Jezik: Slovenski jezik
• Ključne besede: zaznavanje stresa rastlin, spektralno slikanje, umetna inteligenca, strojno učenje, fenotipizacija rastlin
• Vrsta gradiva: Doktorsko delo/naloga
• Tipologija: 2.08 - Doktorska disertacija
• Organizacija: FS - Fakulteta za strojništvo
• Leto izida: 2025
• Št. strani: XXII, 67 str., [72] f. pril.
• PID: 20.500.12556/RUL-175815
• UDK: 004.85:004.9:543.42(043.3)
• COBISS.SI-ID: 256723971
• Datum objave v RUL: 09.11.2025

Sekundarni jezik

• Jezik: Angleški jezik
• Naslov: Early detection of biotic and abiotic stress of potato (Solanum tuberosum L.) plants using spectral imaging and artificial intelligence

Izvleček

Early detection of plant health is crucial for ensuring crop health, optimizing yields, and supporting sustainable agricultural practices. The timely identification of stress factors is enabled through spectral imaging, which has emerged as a powerful, non-invasive tool for monitoring plant stress responses. This thesis investigates the integration of spectral imaging and machine learning techniques for detecting various stressors in potato plants, utilizing multispectral and hyperspectral data collected under both controlled and field conditions. To enhance stress detection accuracy and interpretability, advanced modeling approaches were employed, including deep neural networks, boosting models, and explainability techniques. Feature selection methods were further applied to identify key spectral wavelengths associated with specific stress responses, contributing to a deeper understanding of plant physiological changes under varying stress conditions. The results demonstrate that machine learning models can effectively differentiate between healthy and stressed plants, even before visible symptoms appear. Notably, these models achieved an F1 score of up to 0.95 for identifying water-stressed plants, 100% reliability in detecting nematode-infested potato plants and tubers, and an F1 score of 0.826 for detection of early blight infections. However, the findings also highlight the challenges of distinguishing between biotic and abiotic stressors, particularly in cases of concurrent stress factors. The analysis identified key spectral bands for stress detection: 475–580 nm, 660–730 nm, 940–970 nm, 1420–1510 nm, 1875–2040 nm, and 2350–2480 nm (water stress in potato plants). The near-infrared part of the spectrum was identified as significant for early blight detection, followed by blue and green bands. Additionally, 1000–1200 nm was linked to the insides of nematode-infested tubers, while 1500–1600 nm, 1850–2000 nm, and 2300–2450 nm corresponded to their outsides. These findings provide valuable insights into plant physiology and have significant implications for precision agriculture, offering advancements in disease detection. By integrating spectral imaging with artificial intelligence, this research contributes to the development of innovative tools for mitigating the challenges posed by climate change and plant pests and diseases.

• Ključne besede: plant stress detection, spectral imaging, artificial intelligence, machine learning, plant phenotyping